A Unified Bayesian Approach to Extract Network-Based Functional Differences from a Heterogeneous Patient Cohort

Abstract

We present a generative Bayesian framework that automatically extracts the hubs of altered functional connectivity between a neurotypical and a patient group, while simultaneously incorporating an observed clinical severity measure for each patient. The key to our framework is the latent or hidden organization in the brain that we cannot directly access. Instead, we observe noisy measurements of the latent structure through functional connectivity data. We derive a variational EM algorithm to infer both the latent network topology and the unknown model parameters. We demonstrate the robustness and clinical relevance of our model on a population study of autism acquired at the Kennedy Krieger Institute in Baltimore, MD. Our model results implicate a more diverse pattern of functional differences than two baseline techniques, which do not incorporate patient heterogeneity.